US20090268098A1 - Method for detection of scene changes in a video picture sequence - Google Patents

Abstract

The method comprises the following steps:

• • detection of non-static pixels in a picture, with respect to a preceding picture,
  • calculation of the number of non-static pixels in the picture,
  • calculation of the variation, in absolute value, of the number of non-static pixels for the picture, corresponding to the difference between the number of non-static pixels of the picture and that of the preceding picture,
  • detection of a peak in the variation, the picture relative to the variation leading to the peak then being the first picture of the new scene.

The applications relate to video coding, for example video stream switches.

Description

FIELD OF THE INVENTION
• The present invention relates to the domain of coding of video pictures and concerns a method for detection of scene changes in video picture sequences, better known as “cuts”.
DESCRIPTION OF THE PRIOR ART
• The prior art knows methods for detection of scene changes or shot changes exploiting the temporal correlation. The correlation is based for example on the sum of differences in the luminance of pixels of the same coordinates, between two successive pictures in the picture sequence. However, a movement in the sequence risks being interpreted as a shot change. One solution to this problem consists in establishing a histogram on the picture. This histogram consists in the measurement of the number of pixels or the number of occurrences in the picture for each luminance value. The marker is defined by the luminance values, for example 0 to 255; on the abscissas axis and by the number of occurrences on the ordinate axis. The curve obtained is compared with that of the following picture and their similarity enables declaring a shot change or not. This solution however is not ideal, particularly for erratic movements, or the disappearance or appearance of objects in the scene leading to the appearance or occlusion of pixels behind the objects.
• The exploitation of motion vector fields, for example the non-continuity of motion from one picture to another, does not enable reliable detection due to the fact that the motion vectors are not always representative of physical motion, in particular for motion estimation by “block matching”, based on the correlation rate of a block of a current picture with blocks of a search window of a previous picture.
SUMMARY OF THE INVENTION
• One of the purposes of the invention is to overcome the aforementioned disadvantages. The purpose of the invention is a method for the detection of changes of scene in a video picture sequence, characterised in that it comprises the following steps:
• • detection of non-static pixels in a picture, with respect to a preceding picture,
  • calculation of the number of non-static pixels in the picture,
  • calculation of the variation, in absolute value, of the number of non-static pixels for the picture, corresponding to the difference between the number of non-static pixels of the picture and that of the preceding picture,
  • detection of a peak in the variation, the picture relative to the variation leading to the peak then being the first picture of the new scene.
• According to a particular implementation, the peak detection step comprises a calculation step of the difference of the variation for a picture with respect to at least one preceding picture and with respect to at least one following picture, the detection being obtained if the differences are greater than a predetermined threshold.
• According to a particular implementation, the calculations of difference are carried out for pictures noted as T-(n−1) to T found inside a temporal sliding window bounding a predefined number n of pictures, a peak being declared when the difference, for a picture at a predefined position in the window and for the following picture is greater than a threshold, for the set of pictures in the window other than these two pictures.
• According to a particular implementation, the number n is equal to 8, the pictures of the window being pictures noted as T-7 to T, the two pictures being the pictures noted as T-3 and T-2.
• According to a particular implementation, the picture at a predefined position is the picture at position T-(n/2−1), a picture mean luminance is calculated for the n/2 first pictures of the window and for the n/2 last pictures of the window, these mean luminance values are compared and the threshold is a function of this comparison result.
• According to a particular implementation, the detection step of static areas comprises a comparison of a block of pixels of a current picture with a co-localised block of a preceding picture, the central pixel of the block of the current picture being declared static if the sum of absolute values of pixel luminance differences relating to the neighbours of the central pixel is less than a predetermined threshold.
• According to a particular implementation, the block of pixels is a block of 3×3 pixels and the pixel at the centre is declared static if the sum of absolute values of differences carried out for the pixels around the pixel at the centre is less than a predetermined threshold.
• According to a particular implementation, the video picture sequence is a sequence of interlaced pictures and the detection of static areas is carried out on fields of the same parity.
• The method performs a detection of static areas in the picture, static areas with respect to a preceding picture. A calculation of the number of pixels not belonging to the static area is carried out for each picture then of the variation of this number from one picture to another. Detection of a peak in this variation according to time, that is an increase or decrease, beyond a certain threshold, of the number of pixels of non-static areas, studied over several pictures of the sequence, enables determination of a shot change. One advantage of this solution is to overcome motion in the scene and therefore to provide more reliable results.
• This method is based on the fact that, during a scene change, generally there are not found the same number of pixels in the static or non-static areas.
BRIEF DESCRIPTION OF THE DRAWINGS
• Other specific features and advantages will emerge clearly from the following description, the description provided as a non-restrictive example and referring to the annexed drawings wherein:
• FIG. 1 shows, a detection circuit according to the invention,
• FIG. 2 shows, correlation windows of a picture T and T-1,
• FIG. 3 shows, the variation in time of the number of non-static pixels.
DETAILED DESCRIPTION OF THE EMBODIMENTS OF THE INVENTION
• FIG. 1 represents diagrammatically, the different elements of a circuit implementing the method. The video picture sequence considered is, in the example, a succession of progressive pictures. The video input of the circuit, that receives the video pictures, is connected to a picture delay circuit referenced as1 that memorises a picture of the sequence to transmit it with a picture delay on a second non-static pixeldetection circuit input2. The video input is also connected directly to the second non-static pixeldetection circuit input2. This circuit carries out a detection of non-static pixels in the current picture, detection that is performed between the current video picture at instant T and the preceding picture of the sequence at instant T-1.
• FIG. 2 symbolizes in dotted lines a current picture referenced21 at the instant T and a preceding picture referenced22 at the instant T-1. The processing of these pictures is carried out using windows with dimensions of 3×3 pixels, the pixel at the centre of the window being the current pixel P. For awindow23 ofpicture21, aco-located window24 ofpicture22, that is, of the same emplacement or coordinates, is considered. The sum of absolute values of the luminance differences between these two windows, pixel by pixel, is performed for the 8 pixels around the central pixel. This sum or SAD (Sum of Absolute Differences) is divided by 8 to provide a mean value for the current picture T. If this value is less than a predetermined threshold S1, the value transmitted on a third input of the non-staticpixel detection circuit2, the current point P is labelled as belonging to a static area, as a static pixel. In the contrary case, it is considered to belong to a non-static area, as a non-static pixel. This operation is carried out on each of the pixels of the current picture. A “static” or “non-static” label is thus attributed to each of the pixels of the picture. The non-static pixel or PNS information corresponds for example to the binary value unit and the static pixel information to the binary value zero. The output of the non-staticpixel detection circuit2 is connected to the input of anaccumulator3 to transmit this binary information to it. This latter receives, on a second input, a picture synchronisation or picture reset to zero signal, enabling to reset to zero the content of the accumulator when processing a new picture. Thus, in accordance with the processing of current picture pixels, an accumulation of the number of pixels declared non-static is carried out by this accumulator to provide, in the processing of the picture, a number of non-static pixels NPNS for the complete picture.
• The information relating to the number of non-static pixels in the current picture T exiting theaccumulator3, NPNS_T, is transmitted on afirst register input4 and on afirst subtracter5. The second input ofregister4 receives a picture synchronisation signal that memorises in the register, at each new picture, the NPNS_T information. The memorised information of the preceding picture is, during this recording, extracted from the register to be transmitted on a second input of thesubtracter5. Hence, at the end of processing of the picture T, the subtracter receives on its first input the NPNS_T information from the accumulator and on its second input the NPNS_T-1 information relating to the preceding picture from the register. The signal at the output of the subtracter corresponds to the variation of the number of non-static pixels, of the picture T-1 to picture T, attributed to the picture T and called VAR_NPNS_T or more simply V(T).
• This signal is transmitted to a first register of a succession of 8 elementary registers or memories connected in series to form an 8 cell offset register referenced6, triggered by the picture clock. At the output of each elementary register the VAR_NPNS information is thus available for the pictures T-7 to T. This variation of the number of non-static pixels information on the 8 pictures is transmitted to a comparison and decision circuit referenced7.
• FIG. 3 shows a set of timing diagrams relating to a succession of pictures of the video picture sequence. On the first line is represented a sequence of pictures of the picture T-9 up to the current picture T. The second line represents the NPNS number of non-static pixels for each picture, from the picture T-9 to picture T, the third line shows the unsigned variation of this number, VAR_NPNS, from one picture to another, for these pictures, which is in fact the absolute value of the NPNS gradient from one picture to another. The last line defines a first shot or sequence A up to the picture T-4 and a second shot or sequence B from the picture T-3.
• 
  V(T)=VAR_NPNS_—T=|grad(NPNS)|
• In the example represented in thisFIG. 3, the number of non-static pixels changes slightly from the picture T-9 to the picture T-4 that corresponds to a same sequence A then from the picture T-2 to T which corresponds to a same sequence B. This NPNS number changes greatly for the picture T-3, translating into a shot change between the picture T-4 and the picture T-3, pictures that are taken into account for the calculation of this NPNS number of the picture T-3. It is noted that if for example the sequence A provides a small number of static pixels and if by chance the sequence B provide, in the sequence, this same number, the chances are that their static pixels of the corresponding second sequence, concerning their position in the picture, to those of the first sequence are very low and these static pixels will therefore be considered as non-static for the first picture T-3 of the sequence B.
• The change in the variation of the number of non-static pixels in the sequence translates therefore to a high value for the signal V(T-3), that represents the variation in the number of non-static pixels between the picture T-4 and the picture T-3, attributed to T-3, and a high value for signal V(T-2), that represents the variation in the number of non-static pixels, between the picture T-3 and the picture T-2, attributed to T-2. The signal attributed to the picture T-1, V(T-1), again falls to a low value, NPNS(T-2) being little different to NPNS(T-1). A peak in the signal VAR_NPNS, during a scene change, is therefore noted, this signal being close to zero before the scene change and returning to a value close to zero after the scene change. This is due to the fact that it concerns a signal variation.
• The decision andcomparison circuit7 that receives the signals V(T) to V(T-7) performs a comparison of signal V(T-3) with V(T-7), V(T-6), V(T-5) and V(T-4) then with V(T-1), V(T). It also performs a comparison of the signal V(T-2) with these same values. If the differences corresponding to these 12 comparisons are all greater than a predetermined threshold S2, this signifies that a peak in the variation of the number of non-static pixels was detected, distributed over the pictures T-3 and T-2, and that there is therefore a scene change between the pictures T-4 and T-3. The 3^rdline ofFIG. 3 clearly shows that if the 12 differences are greater than a threshold, this is translated by a value of V(T-3) and of V(T-2) greater than at least this threshold at each of the values V(T) of the other pictures, which corresponds to a difference in the number of non-static pixels, for the picture T-3, with respect to the other pictures.
• In the example, the processing window considered is of 8 pictures, ranging from T-7 to T, the probability of shot change being calculated at picture level T-4 and T-3, enabling taking into account of the scenario prior to and subsequent to this hypothetical shot change. It is of course also conceivable, without leaving the domain of the invention, to consider windows of different sizes with shot change positions in the window, for the calculation of the differences, at different points.
• The use of variation information rather than that of non-static pixel numbers enables display of the value of non-static pixel numbers, that is, to a certain extent, of motion. A homogenous motion in a sequence, whatever the value found of the number of non-static pixels, will provide a variation information of low value.
• An improvement of this method is proposed below and concerns the circuits referenced8 to13 ofFIG. 1.
• The detection of non-static pixels is more delicate for sequences with little luminosity. In fact, the ranges of variation of luminance values being low, it is more delicate to distinguish the static areas from non-static areas by luminance correlation of windows. To support the shot change decision, to render it more reliable, the method implements an analysis of luminance on the pictures of the processing window.
• The video source at the input of the device is transmitted to a picture luminance values extraction circuit referenced8 then to aaccumulator9 that performs the sum of these luminance values on the picture, a picture reset input enabling the accumulator to be emptied for the processing of the following picture. This input is fed by the picture synchronisation signal. The information output from theaccumulator9 is transmitted to a succession of 8 elementary registers in series forming an 8 cell offsetregister10. The picture synchronisation signals enable the transfer of this information from one register to another. The outputs of the 4 first elementary registers containing the luminance information for the pictures T-3 to T are transmitted to an averagingcircuit11 that provides the mean luminance values on these 4 pictures. The outputs of the 4 last elementary registers containing the luminance information for the pictures T-4 to T-7 are transmitted to an averagingcircuit12 that provides the mean luminance values on these 4 pictures.
• Referring toFIG. 3, it then concerns, for thesesignals exiting circuits11 and12, a luminance information that is the mean relative to the 4 pictures of the sequence A, prior to the shot change that took place, in the calculation hypothesis, from picture T-3, and a luminance information that is the mean relative to the 4 pictures of the sequence B, subsequent to the shot change.
• The output of the two averagingcircuits11 and12 are transmitted to asubtracter13 that subtracts one value from the other to transmit the result to a decision andcomparison circuit input7. This mean luminance difference information is used by the decision and comparison circuit to define the thresholds. If the mean luminance values are indeed different, this supports the shot change decision. If the mean luminance values are close, the threshold is modified. It concerns the threshold S2 relative to the 12 comparisons enabling the detection of a peak, value transmitted on an input to the decision andcomparison circuit7 and thus modified by this circuit according to this mean luminance difference.
• It is of course conceivable to realise this mean on different numbers of pictures. The pictures taken into account, for a mean, must be exclusively, either before or after the tested border, that separating the pictures T-4 and T-3 in the example ofFIG. 3.
• In the case where the video sequence is constituted of a succession of interlaced fields, the delayed pictures are replaced by two field delays in order to effect a correlation between two fields of the same parity. The detection of non-static pixels is then made between fields of the same parity and not between successive pictures (frames) in the case of a progressive sequence. The modifications made to the diagram ofFIG. 1 are understood by those skilled in the art and are therefore not detailed here.
• The applications relate to video coding, for example video stream switches.

Claims (8)

1. Method for the detection of changes of scene in a video picture sequence, comprising the following steps:

detection of non-static pixels in a picture, with respect to a preceding picture,

calculation of the number of non-static pixels in the picture,

calculation of the variation, in absolute value, of the number of non-static pixels for the picture, corresponding to the difference between the number of non-static pixels of the picture and that of the preceding picture,

detection of a peak in the variation, the picture relative to the variation leading to the peak then being the first picture of the new scene.

2. Method according toclaim 1, wherein the peak detection step comprises a calculation step of the difference of the variation for a picture with respect to at least one preceding picture and with respect to at least one following picture, the detection being obtained if the differences are greater than a predetermined threshold.

3. Method according toclaim 2, wherein the calculations of difference are carried out for pictures noted as T-(n−1) to T found inside a temporal sliding window bounding a predefined number of n pictures, a peak being declared when the difference, for a picture at a predefined position in the window and for the following picture is greater than a threshold, for the set of pictures in the window other than these two pictures.

4. Method according toclaim 3, wherein the number n is equal to 8, the pictures of the window being pictures noted as T-7 to T, the two pictures being the pictures noted as T-3 and T-2.

5. Method according toclaim 3, wherein the picture at a predefined position is the picture at position T-(n/2−1), in that a picture mean luminance is calculated for the n/2 first pictures of the window and for the n/2 last pictures of the window, wherein these mean luminance values are compared and in that the threshold is a function of this comparison result.

6. Method according toclaim 1, wherein the detection step of static areas comprises a comparison of a block of pixels of a current picture with a co-localised block of a preceding picture and wherein the central pixel of the block of the current picture is declared static if the sum of absolute values of pixel luminance differences relating to the neighbours of the central pixel is less than a predetermined threshold.

7. Method according toclaim 6, wherein the block of pixels is a block of 3×3 pixels and the pixel at the centre is declared static if the sum of absolute values of differences carried out for the pixels around the pixel at the centre is less than a predetermined threshold.

8. Method according toclaim 1, wherein the video picture sequence is a sequence of interlaced pictures and the detection of static areas is carried out on fields of the same parity.

Images